A deep learning-based high-speed railway subgrade vibration anomaly grading early warning method

Abstract

The present application relates to the technical field of rail transit safety intelligent monitoring, in particular to a high-speed railway subgrade vibration anomaly grading early warning method based on deep learning, which deeply integrates the subgrade dynamics priori and deep learning, specifically extracts multi-frequency band anomaly features conforming to physical laws through multi-scale sparse convolution, adopts a working condition adaptive gating mechanism to strip working condition interference such as train speed and axle load, uses dual-domain attention to focus on key anomaly periods and channels, simultaneously outputs risk index and reliability coefficient, and realizes accurate grading early warning in combination with dynamic three-level threshold and adaptive sliding window. The present application can improve the accuracy, stability and decision-making reliability of high-speed railway subgrade vibration anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for rail transit safety, and in particular to a method for graded early warning of abnormal vibrations in high-speed railway subgrades based on deep learning. Background Technology

[0002] With the increase in the operating mileage of high-speed railways and the continuous improvement in train speed and density, the subgrade structure is subjected to cyclic dynamic loads for a long time, making it prone to problems such as settlement, loosening of fill material, and frost heave and thaw settlement, which seriously threaten traffic safety. Vibration monitoring, as a core means of sensing the health status of the subgrade, has been widely used in the high-speed railway operation and maintenance system.

[0003] Existing roadbed vibration anomaly early warning technologies mainly fall into two categories: The first is the traditional method based on fixed thresholds, which triggers an alarm by setting fixed thresholds for vibration amplitude, dominant frequency, and other indicators. While this method is simple in principle, it cannot adapt to dynamic changes in operating conditions such as train speed, axle load, ambient temperature, and track irregularities. In high-speed, heavy-load scenarios, the false alarm rate is as high as 15%, and the missed detection rate for early, slow degradation is also relatively high. The second category is intelligent methods based on machine learning, such as support vector machines and backpropagation neural networks, which classify data by manually extracting time-frequency domain features. However, manual features rely on expert experience, have limited expressive power, and are difficult to capture complex nonlinear degradation patterns.

[0004] In recent years, deep learning technology has been applied to vibration signal analysis, with models such as convolutional neural networks and long short-term memory networks used for roadbed anomaly identification. However, existing deep learning methods are mostly data-driven, lacking prior constraints on roadbed dynamics, resulting in poor model interpretability and insufficient generalization ability across geological regions. They also fail to effectively isolate the interference of operating conditions on vibration amplitude, leading to persistent false alarms. Furthermore, the warning judgment uses a fixed-length sliding window, making it impossible to dynamically adjust sensitivity based on risk changes. Additionally, they only output binary or hierarchical results, failing to quantify inference uncertainties, thus posing risks to operation and maintenance decisions. Therefore, there is an urgent need for a hierarchical early warning method for high-speed railway roadbed vibration anomalies that integrates physical priors, adapts to changes in operating conditions, and possesses reliability assessment capabilities.

[0005] Therefore, this invention proposes a deep learning-based method for graded early warning of vibration anomalies in high-speed railway subgrades to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a deep learning-based method for hierarchical early warning of vibration anomalies in high-speed railway subgrades. This invention can improve the accuracy, stability, and decision reliability of vibration anomaly identification in high-speed railway subgrades.

[0007] The technical solution of this invention to solve the technical problem is a deep learning-based method for graded early warning of vibration anomalies in high-speed railway subgrades, comprising the following steps: S1. Deploy an array of acceleration sensors along the high-speed railway line to collect the original three-axis vibration signals of the roadbed and the auxiliary data of synchronous working conditions when the train passes. Preprocess the original three-axis vibration signals and output a standardized multi-channel vibration time sequence, synchronous working condition feature vector and roadbed reference dynamic parameter matrix. S2. Construct a multi-scale sparse convolution feature extraction module based on the prior knowledge of subgrade dynamics. With the subgrade vibration transmission elastic half-space equation as a constraint, multi-band abnormal features are extracted in parallel by three-branch frequency band matching sparse convolution kernel groups. Multi-scale sparse local feature maps are output through cascaded residual fusion. S3. Construct a working condition adaptive gated bidirectional long short-term memory network time-series module. Input the local feature map and working condition feature vector, remove the interference of the feature by working condition normalization gate at each time step, and then model the long-term time-series dependency through a two-layer bidirectional LSTM to output the roadbed anomaly time-series representation. S4. Construct a physical prior-guided wheelset-sensitive dual-domain attention module. In the channel dimension, combine vibration-sensitive priors to set initial weights, and in the temporal dimension, combine wheelset loading priors to set position biases. Output anomaly discrimination fusion features through joint weighting. S5. Construct a fully connected quantization layer with physical constraints, input the anomaly discrimination fusion features into this layer, combine it with the roadbed benchmark dynamic parameter matrix, and output the risk index and risk reliability system. S6. Calculate the dynamic three-level risk threshold based on the working condition parameters, perform initial classification of the risk index, construct an adaptive sliding window to cache the results of N consecutive periods, adjust the window length according to the risk change trend, and update the final warning status only when the samples in the window meet the level consistency and trend constraints. S7. Implement tiered closed-loop handling based on the warning level. Edge nodes regularly upload inference data and operation and maintenance feedback to the cloud. The system achieves self-iteration by fine-tuning the model through incremental learning and issuing updates via OTA.

[0008] S1 is as follows: S1.1, Operating condition auxiliary data includes train operating condition information, track geometric parameters, roadbed reference dynamic parameters, and environmental meteorological data; A single monitoring cycle is defined as the complete duration from when the front end of the train enters the sensor's sensing area to when the rear end completely leaves. A uniformly aligned duration L is set, and the corresponding time sequence length is L. Train operating information includes train type, axle load, operating speed, and passing time; track geometry parameters include elevation, alignment, levelness, and triangular embankment; subgrade reference dynamic parameters include dynamic stiffness, damping ratio, and allowable dynamic stress; and environmental meteorological data includes temperature, humidity, and rainfall. S1.2. The baseline of the triaxial original vibration signal is corrected by using third-order polynomial fitting. Then, the peak detection algorithm is used to extract the wheel pair passing feature pulse in the vertical vibration signal. Taking the first wheel pair pulse as the time zero point, a fixed time sequence of length L is extracted to complete the wheel pair reference alignment and obtain the aligned vibration sequence. S1.3. Using the detection cycle as the unit, the working condition auxiliary data and the aligned vibration sequence are timestamped to construct a synchronous working condition feature vector. Then, the roadbed reference dynamic parameter matrix is ​​constructed with the three-axis vibration channel as the row and the roadbed reference dynamic parameter as the column. Then, the vibration sequence and the synchronous working condition feature vector are Z-score standardized to obtain a standardized multi-channel vibration time sequence.

[0009] S2 is as follows: The multi-scale sparse convolution feature extraction module based on roadbed dynamics priors contains four cascaded multi-scale sparse convolution units with kernel lengths of 3, 15, and 31, respectively, which capture high-frequency local damage features, mid-frequency wheelset inherent vibration features, and low-frequency overall roadbed deformation features. Each branch road is equipped with sparse regular constraints based on roadbed dynamics priors, and each unit is equipped with identity residual connections. The number of output channels of the convolution kernels of the four multi-scale sparse convolution units are 32, 64, 128 and 256 respectively. After each sparse convolution operation, a batch normalization layer, a ReLU activation function layer and a max pooling layer are connected in sequence. The pooling window size is 3 and the stride is 2. Finally, the multi-scale sparse local feature map is output.

[0010] S3 is as follows: The timing module of the adaptive gated bidirectional long short-term memory network has a two-layer stacked structure, with each layer containing forward LSTM units and backward LSTM units, and the number of hidden units in one direction is 256. The synchronous operating condition feature vector is mapped to a 256-dimensional operating condition gating feature through two fully connected layers, and then the amplitude of the multi-scale sparse local features is normalized through a time-step operating condition normalization gating mechanism. Then, the normalized temporal features are input into a two-layer bidirectional LSTM network. The forward network captures the evolution of vibration signals along the forward temporal direction, and the backward network captures contextual dependencies along the reverse temporal direction. The bidirectional captured features are fused to output a temporal representation of roadbed anomalies.

[0011] S4 is as follows: The physics-prior-guided wheelset-sensitive dual-domain attention module consists of two parts: channel attention and temporal attention. The two units are computed in parallel and then jointly weighted element by element. The channel attention calculation is based on the roadbed reference dynamic parameter matrix setting to set the physical prior of roadbed vibration, and sets the initial weights of the three channels respectively. The roadbed anomaly time series characterization is input, and the adaptive channel weights are calculated through global average pooling and fully connected layer. Temporal attention calculation is based on the physical prior of the wheelset passage time period, and sets the temporal position corresponding to the wheelset pulse. The weighted bias is doubled, and the anomaly discrimination weights of the roadbed anomaly time series representation at each time step are calculated by dot product attention. The adaptive channel weights and anomaly detection weights are jointly weighted to output a weighted fusion feature.

[0012] S5 is detailed below: The physically constrained fully connected quantization layer consists of two fully connected layers. The first layer maps the anomaly detection fusion features into a 128-dimensional vector, and the second layer outputs both the risk index R and the risk reliability coefficient K. The loss function of the fully connected layer adds a roadbed dynamics constraint term to the weighted binary cross-entropy loss, and the risk reliability coefficient K is obtained by quantifying the uncertainty of model inference using the Monte Carlo dropout method, K∈[0,1]. When K < 0.5, data re-sampling and secondary inference are automatically triggered. If K is still less than 0.5 after secondary inference, the system is marked as awaiting manual review.

[0013] S6 is detailed below: The formula for calculating the adaptive dynamic level 3 risk threshold is as follows: , , , in, This indicates the deviation between the current train speed and the reference speed. This indicates the deviation between the current axle load and the reference axle load. This indicates the deviation between the current ambient temperature and the reference temperature. This indicates the deviation of the current orbital geometric irregularity index from the baseline value; The classification rules are as follows: like This is considered normal. like This triggered a yellow alert. like This triggered an orange alert. like This triggered a red alert. Build length is The sliding window cache recently Risk index and initial warning level for each monitoring period, window length The adaptive adjustment rule is: For high-speed railway trunk lines designed for a speed of 350 km / h, The value is 5; For passenger dedicated lines designed for a speed of 250 km / h, The value is 4; For other routes The value is 3; When the risk index rises for three consecutive cycles, the slope of the rise is... hour, Rounded down to 60% of the benchmark value; when the risk index volatility coefficient within the window... hour, Round up to 150% of the base value, with a maximum of 8; The final warning status is updated only when the initial warning level of all samples within the window is consistent and the trend of risk index changes matches the warning level; otherwise, the original warning status is maintained and a log is generated, recording it as a status conflict.

[0014] S7 is detailed below: Yellow alert: Automatically generate inspection task orders containing monitoring point locations, risk indices, and reliability coefficients, and push them to the engineering maintenance platform; Orange alert: Automatically sends speed limit instructions to the train dispatching system and generates a special inspection task order; Red alert: Automatically sends an emergency braking command to the train dispatching system, closes the relevant section, and initiates emergency repair procedures; The early warning message is encapsulated in JSON format, including the geographical coordinates of the monitoring point, risk index R, reliability coefficient K, warning level, trigger time, associated train number, and suggested measures, and is simultaneously pushed to the track maintenance platform, train dispatching system, and GIS platform.

[0015] The specific deployment of the accelerometer array is as follows: An accelerometer array is arranged at fixed intervals on both sides of the track shoulder. Each monitoring point is equipped with a triaxial MEMS accelerometer. The triaxial MEMS accelerometer has a range of ±2g, a resolution greater than 50μg, and an operating temperature range of -40℃ to +85℃. Each accelerometer integrates a BeiDou time synchronization module for synchronizing the time across the entire line.

[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a hierarchical early warning method for vibration anomalies in high-speed railway subgrades based on a spatiotemporal evolution fusion network. By deeply fusing physical priors and deep learning, the elastic half-space equation for subgrade vibration transmission is introduced as a constraint into the convolutional layer. Sparse regularization guides the model to learn features conforming to physical laws, improving generalization and interpretability, and avoiding the overfitting problem of purely data-driven models. This invention employs a condition-adaptive gating mechanism, dynamically eliminating the influence of train speed, axle load, temperature, and track irregularities on vibration amplitude through time-step condition normalization, thereby effectively eliminating condition interference. The sensitive priors for subgrade vibration channels and wheelset loading periods guide the model to focus on specific areas. By focusing on the vertical passage and wheelset interaction periods, which best reflect roadbed anomalies, the accuracy of anomaly identification can be improved. This invention introduces a risk reliability coefficient to quantify inference uncertainty. The reliability coefficient is calculated using the Monte Carlo dropout method, triggering secondary inference or manual review for low-confidence results, thereby reducing decision-making risk. This invention also dynamically adjusts the warning threshold based on real-time operating conditions and adaptively adjusts the sliding window length in conjunction with risk change trends, which can reduce the number of warning state switching while ensuring timely warnings. Furthermore, this invention shortens the update cycle by periodically uploading anonymized data through edge nodes, fine-tuning the model based on incremental learning in the cloud, and remotely distributing updates. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0019] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0020] Example 1 like Figure 1 As shown, a deep learning-based method for graded early warning of vibration anomalies in high-speed railway subgrade includes the following steps: S1. Deploy an array of acceleration sensors along the high-speed railway line to collect the original three-axis vibration signals of the roadbed and the auxiliary data of synchronous working conditions when the train passes. Preprocess the original three-axis vibration signals and output a standardized multi-channel vibration time sequence, synchronous working condition feature vector and roadbed reference dynamic parameter matrix. S2. Construct a multi-scale sparse convolution feature extraction module based on the prior knowledge of subgrade dynamics. With the subgrade vibration transmission elastic half-space equation as a constraint, multi-band abnormal features are extracted in parallel by three-branch frequency band matching sparse convolution kernel groups. Multi-scale sparse local feature maps are output through cascaded residual fusion. S3. Construct a working condition adaptive gated bidirectional long short-term memory network time-series module. Input the local feature map and working condition feature vector, remove the interference of the feature by working condition normalization gate at each time step, and then model the long-term time-series dependency through a two-layer bidirectional LSTM to output the roadbed anomaly time-series representation. S4. Construct a physical prior-guided wheelset-sensitive dual-domain attention module. In the channel dimension, combine vibration-sensitive priors to set initial weights, and in the temporal dimension, combine wheelset loading priors to set position biases. Output anomaly discrimination fusion features through joint weighting. S5. Construct a fully connected quantization layer with physical constraints, input the anomaly discrimination fusion features into this layer, combine it with the roadbed benchmark dynamic parameter matrix, and output the risk index and risk reliability system. S6. Calculate the dynamic three-level risk threshold based on the working condition parameters, perform initial classification of the risk index, construct an adaptive sliding window to cache the results of N consecutive periods, adjust the window length according to the risk change trend, and update the final warning status only when the samples in the window meet the level consistency and trend constraints. S7. Implement tiered closed-loop handling based on the warning level. Edge nodes regularly upload inference data and operation and maintenance feedback to the cloud. The system achieves self-iteration by fine-tuning the model through incremental learning and issuing updates via OTA.

[0021] In a specific implementation, S1 is as follows: S1.1, Operating condition auxiliary data includes train operating condition information, track geometric parameters, roadbed reference dynamic parameters, and environmental meteorological data; A single monitoring cycle is defined as the complete duration from when the front end of the train enters the sensor's sensing area to when the rear end completely leaves. A uniformly aligned duration L is set, and the corresponding time sequence length is L. Train operating information includes train type, axle load, operating speed, and passing time; track geometry parameters include elevation, alignment, levelness, and triangular embankment; subgrade reference dynamic parameters include dynamic stiffness, damping ratio, and allowable dynamic stress; and environmental meteorological data includes temperature, humidity, and rainfall. S1.2. The baseline of the triaxial original vibration signal is corrected by using third-order polynomial fitting. Then, the peak detection algorithm is used to extract the wheel pair passing feature pulse in the vertical vibration signal. Taking the first wheel pair pulse as the time zero point, a fixed time sequence of length L is extracted to complete the wheel pair reference alignment and obtain the aligned vibration sequence. S1.3. Using the detection cycle as the unit, the working condition auxiliary data and the aligned vibration sequence are timestamped to construct a synchronous working condition feature vector. Then, the roadbed reference dynamic parameter matrix is ​​constructed with the three-axis vibration channel as the row and the roadbed reference dynamic parameter as the column. Then, the vibration sequence and the synchronous working condition feature vector are Z-score standardized to obtain a standardized multi-channel vibration time sequence.

[0022] In a specific implementation, the deployed acceleration sensor array is as follows: An accelerometer array is arranged at fixed intervals on both sides of the track shoulder. Each monitoring point is equipped with a triaxial MEMS accelerometer. The triaxial MEMS accelerometer has a range of ±2g, a resolution greater than 50μg, and an operating temperature range of -40℃ to +85℃. Each accelerometer integrates a BeiDou time synchronization module for synchronizing the time across the entire line.

[0023] In a specific implementation, S2 is as follows: The multi-scale sparse convolution feature extraction module based on roadbed dynamics priors contains four cascaded multi-scale sparse convolution units with kernel lengths of 3, 15, and 31, respectively, which capture high-frequency local damage features, mid-frequency wheelset inherent vibration features, and low-frequency overall roadbed deformation features. Each branch road is equipped with sparse regular constraints based on roadbed dynamics priors, and each unit is equipped with identity residual connections. The kernel weight constraint formula for sparse convolution is: , in, Represents a standardized multi-channel vibration time series. Represents the sparse convolution kernel weight matrix. express Take the minimum value. This represents the convolution operation. The feature map representing the output of the sparse convolution branch. This represents the regularity coefficient calculated based on the roadbed reference dynamic parameters. It represents the Hadamah accumulation. This represents the dynamic prior mask matrix, where the weight of the convolution kernel position corresponding to the roadbed vibration sensitive frequency band is 1, and the weight of the non-sensitive frequency band is 0. The mask matrix needs to be fine-tuned according to the roadbed dynamic characteristics of different geological regions. In soft soil areas, the weight of the low-frequency channel needs to be increased. No. The output calculation formula for a multi-scale sparse convolutional unit is as follows: , in, Indicates the first The input of a multi-scale sparse convolutional unit, Indicates the first The kernel length within each multi-scale sparse convolutional unit is... sparse convolution operation, This indicates a splicing operation. The linear mapping weight matrix represents the residual connection. Indicates the first Feature maps output by multi-scale sparse convolutional units; The number of output channels of the convolution kernels of the four multi-scale sparse convolution units are 32, 64, 128 and 256 respectively. After each sparse convolution operation, a batch normalization layer, a ReLU activation function layer and a max pooling layer are connected in sequence. The pooling window size is 3 and the stride is 2. Finally, the multi-scale sparse local feature map is output.

[0024] Multi-scale convolution comprehensively covers the vibration frequency bands corresponding to roadbed anomalies, which can avoid the loss of information from single-scale features. Physical constraints guide the model to learn features that conform to the dynamic laws of roadbed, which can improve the cross-regional generalization ability.

[0025] In a specific implementation, S3 is as follows: The timing module of the adaptive gated bidirectional long short-term memory network has a two-layer stacked structure, with each layer containing forward LSTM units and backward LSTM units, and the number of hidden units in one direction is 256. The synchronous operating condition feature vector is mapped to a 256-dimensional operating condition gating feature through two fully connected layers, and then the amplitude of the multi-scale sparse local features is normalized through a time-step operating condition normalization gating mechanism. The calculation formula is as follows: , in, Indicates the first Local features at each time step express Activation function Indicates the gating characteristics of the operating conditions. Indicates the first Features after normalization at each time step; Then, the normalized temporal features are input into a two-layer bidirectional LSTM network. The forward network captures the evolution of vibration signals along the forward temporal direction, and the backward network captures contextual dependencies along the reverse temporal direction. The bidirectional captured features are fused to output a temporal representation of roadbed anomalies.

[0026] The time-step normalized dynamics eliminates the influence of operating conditions on vibration amplitude, making the abnormal characteristics under different operating conditions comparable; the bidirectional LSTM effectively captures the long-range temporal dependence of vibration signals, improving the ability to identify slow degradation anomalies.

[0027] In a specific implementation, S4 is as follows: The physics-prior-guided wheelset-sensitive dual-domain attention module consists of two parts: channel attention and temporal attention. The two units are computed in parallel and then jointly weighted element by element. The channel attention calculation is based on the roadbed reference dynamic parameter matrix setting to set the physical prior of roadbed vibration, and sets the initial weights of the three channels respectively. The roadbed anomaly time series characterization is input, and the adaptive channel weights are calculated through global average pooling and fully connected layer. Temporal attention calculation is based on the physical prior of the wheelset passage time period, and sets the temporal position corresponding to the wheelset pulse. The weighted bias is doubled, and the anomaly discrimination weights of the roadbed anomaly time series representation at each time step are calculated by dot product attention. The adaptive channel weights and anomaly detection weights are jointly weighted to output a weighted fusion feature. The initial channel weights can be fine-tuned based on actual line monitoring data; the wheelset position offset must strictly correspond to the wheelset pulses aligned in S1. The dual-domain attention-guided model focuses on the channels and time periods that best reflect roadbed anomalies, which can suppress interference from irrelevant information; physical prior initialization accelerates the convergence of the attention module and can improve the training efficiency of the model.

[0028] In a specific implementation, S5 is as follows: The physically constrained fully connected quantization layer consists of two fully connected layers. The first layer maps the anomaly detection fusion features into a 128-dimensional vector, and the second layer outputs both the risk index R and the risk reliability coefficient K. The loss function of the fully connected layer adds a roadbed dynamics constraint term to the weighted binary cross-entropy loss, and the risk reliability coefficient K is obtained by quantifying the uncertainty of model inference using the Monte Carlo dropout method, K∈[0,1]. When K < 0.5, data re-sampling and secondary inference are automatically triggered. If K is still less than 0.5 after secondary inference, the system is marked as awaiting manual review.

[0029] Loss function of fully connected layer The calculation formula is as follows: ， in, This represents the weighted binary cross-entropy loss. Indicates the physical constraint weight coefficient. This represents the theoretical risk value calculated based on the roadbed reference dynamic parameters; Theoretical risk value The calculation formula is as follows: , in, This represents the baseline risk reliability coefficient. This represents the current risk reliability coefficient. Indicates the reference damping ratio. This indicates the current damping ratio.

[0030] The reliability coefficient quantifies the uncertainty of model inference, avoiding incorrect decisions caused by low-confidence results; the physical constraint loss makes the model output more consistent with the laws of roadbed dynamics, improving the interpretability of the results.

[0031] In a specific implementation, S6 is as follows: The formula for calculating the adaptive dynamic level 3 risk threshold is as follows: , , , in, This indicates the deviation between the current train speed and the reference speed. This indicates the deviation between the current axle load and the reference axle load. This indicates the deviation between the current ambient temperature and the reference temperature. This indicates the deviation of the current orbital geometric irregularity index from the baseline value; The classification rules are as follows: like This is considered normal. like This triggered a yellow alert. like This triggered an orange alert. like This triggered a red alert. Build length is The sliding window cache recently Risk index and initial warning level for each monitoring period, window length The adaptive adjustment rule is: For high-speed railway trunk lines designed for a speed of 350 km / h, The value is 5; For passenger dedicated lines designed for a speed of 250 km / h, The value is 4; For other routes The value is 3; When the risk index rises for three consecutive cycles, the slope of the rise is... hour, Rounded down to 60% of the benchmark value; when the risk index volatility coefficient within the window... hour, Round up to 150% of the base value, with a maximum of 8; The formula for calculating the upward slope of the risk index is as follows: , The formula for calculating the volatility coefficient of the risk index is as follows: , in, Indicates the first Risk index for each monitoring period Indicates the monitoring period index, This represents the standard deviation of the risk index within the window. This represents the average risk index within the window.

[0032] The final warning status is updated only when the initial warning level of all samples within the window is consistent and the trend of risk index changes matches the warning level; otherwise, the original warning status is maintained and a log is generated, recording it as a status conflict.

[0033] In a specific implementation, S7 is as follows: Yellow alert: Automatically generate inspection task orders containing monitoring point locations, risk indices, and reliability coefficients, and push them to the engineering maintenance platform; Orange alert: Automatically sends speed limit instructions to the train dispatching system and generates a special inspection task order; Red alert: Automatically sends an emergency braking command to the train dispatching system, closes the relevant section, and initiates emergency repair procedures; The early warning message is encapsulated in JSON format, including the geographical coordinates of the monitoring point, risk index R, reliability coefficient K, warning level, trigger time, associated train number, and suggested measures, and is simultaneously pushed to the track maintenance platform, train dispatching system, and GIS platform.

[0034] The specific steps for model training and edge-cloud collaborative deployment are as follows: (1) Model training: The training is conducted in an end-to-end manner. The training dataset contains no less than 500,000 historical vibration signal samples from different geological regions, different train types, different operating speeds and different ambient temperatures. Among them, the proportion of abnormal samples is no less than 15%. The abnormal samples cover typical patterns such as early settlement, loose fill, external construction disturbance, and frost heave and thaw settlement. The data is augmented to 3 times the original size through time stretching, amplitude scaling and Gaussian noise injection. (2) Edge deployment: The model is deployed on edge computing nodes based on ARM architecture. Each node is responsible for processing sensor data within a 5-kilometer range and is equipped with a dedicated AI acceleration chip. The time for a single inference is less than 50 milliseconds. The nodes are interconnected via gigabit industrial Ethernet. (3) Cloud collaboration: Every 24 hours, the edge nodes upload the anonymized inference results (input features, R value, K value, operation and maintenance feedback) to the cloud data center. The cloud performs online fine-tuning of the model every 7 days based on the incremental learning algorithm. After fine-tuning, the model is distributed to the edge nodes via OTA.

[0035] Example 2 To verify the superiority of the method of this invention, an experimental dataset was constructed using monitoring data from a soft soil section and a karst section of a high-speed railway. The data covers typical anomaly modes such as initial settlement, loose fill material, frost heave and thaw settlement, and external construction disturbance. The comparison methods include: Method A: The method proposed in this invention; Method B: 1D-CNN-BiLSTM-Attention method; Method C: Traditional fixed threshold method; Method D: SVM + manual time-frequency feature method; Experimental environment: Intel Xeon Gold 6248R CPU, NVIDIA RTX 3090 GPU, PyTorch 2.0 framework. Evaluation metrics include accuracy, false alarm rate, false negative rate, average alert response time, and cross-region generalization accuracy. The experimental results are shown in Table 1. Table 1. Experimental Comparison and Evaluation Table As shown in Table 1, the method of this invention achieves an accuracy of 98.7% in soft soil sections, with a false alarm rate of only 0.28%, representing an improvement of 6.4% and a reduction of 86.7% respectively compared to Method B, significantly outperforming existing technologies. Method A achieves a cross-regional generalization accuracy of 96.2%, a 7.7% improvement over Method B, demonstrating that physical prior fusion effectively enhances the model's generalization ability. The number of warning state switching times is only 1.2 times / day, a reduction of 86.2% compared to Method B, indicating that the adaptive sliding window mechanism significantly improves system stability. The average warning response time is 42ms, meeting real-time monitoring requirements and providing sufficient time for emergency response. In summary, this invention demonstrates that the method of this invention can improve the accuracy, stability, and decision-making reliability of high-speed railway subgrade vibration anomaly identification.

[0036] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A deep learning-based high-speed railway track vibration anomaly grading early warning method, characterized in that, Includes the following steps: S1. Deploy an array of acceleration sensors along the high-speed railway line to collect the original three-axis vibration signals of the roadbed and the auxiliary data of synchronous working conditions when the train passes. Preprocess the original three-axis vibration signals and output a standardized multi-channel vibration time sequence, synchronous working condition feature vector and roadbed reference dynamic parameter matrix. S2. Construct a multi-scale sparse convolution feature extraction module based on the prior knowledge of subgrade dynamics. With the subgrade vibration transmission elastic half-space equation as a constraint, multi-band abnormal features are extracted in parallel by three-branch frequency band matching sparse convolution kernel groups. Multi-scale sparse local feature maps are output through cascaded residual fusion. S2 is as follows: The multi-scale sparse convolution feature extraction module based on roadbed dynamics priors contains four cascaded multi-scale sparse convolution units. The kernel lengths in each multi-scale sparse convolution unit are 3, 15, and 31, respectively, which capture high-frequency local damage features, mid-frequency wheelset inherent vibration features, and low-frequency overall roadbed deformation features. Each branch road is equipped with sparse regular constraints based on roadbed dynamics priors, and each unit is equipped with identity residual connections. The number of output channels of the convolution kernels of the four multi-scale sparse convolution units are 32, 64, 128 and 256 respectively. After each sparse convolution operation, a batch normalization layer, a ReLU activation function layer and a max pooling layer are connected in sequence. The pooling window size is 3 and the stride is 2. Finally, the multi-scale sparse local feature map is output. S3. Construct a working condition adaptive gated bidirectional long short-term memory network time-series module. Input the local feature map and working condition feature vector. Remove the interference of working condition on the feature through the working condition normalization gating mechanism at each time step. Then use a two-layer bidirectional LSTM to mine the long-range temporal dependency of the corrected feature and output the roadbed anomaly temporal representation. S3 is as follows: The timing module of the adaptive gated bidirectional long short-term memory network has a two-layer stacked structure, with each layer containing forward LSTM units and backward LSTM units, and the number of hidden units in one direction is 256. The synchronous operating condition feature vector is mapped to a 256-dimensional operating condition gating feature through two fully connected layers. Then, the operating condition normalization gating mechanism is used step-by-step to normalize the amplitude of the multi-scale sparse local feature map. Then, the normalized temporal features are input into a two-layer bidirectional LSTM network. The forward network captures the evolution of vibration signals along the forward temporal direction, and the backward network captures contextual dependencies along the reverse temporal direction. The bidirectional captured features are fused to output a temporal representation of roadbed anomalies. S4. Construct a physical prior-guided wheelset-sensitive dual-domain attention module. In the channel dimension, combine vibration-sensitive priors to set initial weights, and in the temporal dimension, combine wheelset loading priors to set position biases. Output anomaly discrimination fusion features through joint weighting. S5. Construct a fully connected quantization layer with physical constraints, input the anomaly discrimination fusion features into this layer, combine it with the roadbed benchmark dynamic parameter matrix, and output the risk index and risk reliability coefficient. S6. Calculate the dynamic three-level risk threshold based on the working condition parameters, perform initial classification of the risk index, construct an adaptive sliding window to cache the results of N consecutive periods, adjust the window length according to the risk change trend, and update the final warning status only when the samples in the window meet the level consistency and trend constraints. S7. Implement tiered closed-loop handling based on the warning level. Edge nodes regularly upload inference data and operation and maintenance feedback to the cloud. The system achieves self-iteration by fine-tuning the model through incremental learning and issuing updates via OTA.

2. The deep learning-based rail track foundation vibration anomaly grading and early warning method according to claim 1, characterized in that, S1 is as follows: S1.1, Operating condition auxiliary data includes train operating condition information, track geometric parameters, roadbed reference dynamic parameters, and environmental meteorological data; A single monitoring cycle is defined as the complete duration from when the front end of the train enters the sensor's sensing area to when the rear end completely leaves. A uniformly aligned duration L is set, and the corresponding time sequence length is L. Train operating information includes train type, axle load, operating speed, and passing time; track geometry parameters include elevation, alignment, levelness, and triangular embankment; subgrade reference dynamic parameters include dynamic stiffness, damping ratio, and allowable dynamic stress; and environmental meteorological data includes temperature, humidity, and rainfall. S1.

2. The baseline of the triaxial original vibration signal is corrected by using third-order polynomial fitting. Then, the peak detection algorithm is used to extract the wheel pair passing feature pulse in the vertical vibration signal. Taking the first wheel pair pulse as the time zero point, a fixed time sequence of length L is extracted to complete the wheel pair reference alignment and obtain the aligned vibration sequence. S1.

3. Using the detection cycle as the unit, the working condition auxiliary data and the aligned vibration sequence are timestamped to construct a synchronous working condition feature vector. Then, the roadbed reference dynamic parameter matrix is ​​constructed with the three-axis vibration channel as the row and the roadbed reference dynamic parameter as the column. Then, the vibration sequence and the synchronous working condition feature vector are Z-score standardized to obtain a standardized multi-channel vibration time sequence.

3. The deep learning-based rail track foundation vibration anomaly grading and early warning method according to claim 2, characterized in that, S4 is as follows: The physics-prior-guided wheelset-sensitive dual-domain attention module consists of two parts: channel attention and temporal attention. The two units are computed in parallel and then jointly weighted element by element. The channel attention calculation is based on the roadbed reference dynamic parameter matrix, and the roadbed vibration physical prior is set. The initial weights of the three channels are set respectively, the roadbed anomaly time series characterization is input, and the adaptive channel weights are calculated through global average pooling and fully connected layer. Temporal attention calculation is based on the physical prior of the wheelset passage time period, and sets the temporal position corresponding to the wheelset pulse. The weighted bias is doubled, and the anomaly discrimination weights of the roadbed anomaly time series representation at each time step are calculated by dot product attention. The adaptive channel weights and anomaly detection weights are jointly weighted to output a weighted fusion feature.

4. The method for graded early warning of vibration anomalies in high-speed railway subgrade based on deep learning according to claim 3, characterized in that, S5 is detailed below: The physically constrained fully connected quantization layer consists of two fully connected layers. The first layer maps the anomaly detection fusion features into a 128-dimensional vector, and the second layer outputs both the risk index R and the risk reliability coefficient K. The loss function of the fully connected layer is based on the weighted binary cross-entropy loss with the addition of a roadbed dynamics constraint term. The risk reliability coefficient K is obtained by quantifying the uncertainty of model inference using the Monte Carlo dropout method, and K∈[0,1]. When K < 0.5, data re-sampling and secondary inference are automatically triggered. If K is still less than 0.5 after secondary inference, the system is marked as awaiting manual review.

5. The method for graded early warning of vibration anomalies in high-speed railway subgrade based on deep learning according to claim 4, characterized in that, S6 Specifically as follows: The formula for calculating the adaptive dynamic level 3 risk threshold is as follows: ， ， ， in, This indicates the deviation between the current train speed and the reference speed. This indicates the deviation between the current axle load and the reference axle load. This indicates the deviation between the current ambient temperature and the reference temperature. This indicates the deviation of the current orbital geometric irregularity index from the baseline value; The classification rules are as follows: like This is considered normal. like This triggered a yellow alert. like This triggered an orange alert. like This triggered a red alert. Build length is The sliding window cache recently Risk index and initial warning level for each monitoring period, window length The adaptive adjustment rule is: For high-speed railway trunk lines designed for a speed of 350 km / h, The value is 5; For passenger dedicated lines designed for a speed of 250 km / h, The value is 4; For other lines The value is 3; When the risk index rises for three consecutive cycles, the slope of the rise is... hour, Rounded down to 60% of the benchmark value; when the risk index volatility coefficient within the window... hour, Round up to 150% of the base value, with a maximum of 8; The final warning status is updated only when the initial warning level of all samples within the window is consistent and the trend of risk index changes matches the warning level; otherwise, the original warning status is maintained and a log is generated, recording it as a status conflict.

6. The method for graded early warning of vibration anomalies in high-speed railway subgrade based on deep learning according to claim 5, characterized in that, S7 is detailed below: Yellow alert: Automatically generate inspection task orders containing monitoring point locations, risk indices, and reliability coefficients, and push them to the engineering maintenance platform; Orange alert: Automatically sends speed limit instructions to the train dispatching system and generates a special inspection task order; Red alert: Automatically sends an emergency braking command to the train dispatching system, closes the relevant section, and initiates emergency repair procedures; The early warning message is encapsulated in JSON format, including the geographical coordinates of the monitoring point, risk index R, reliability coefficient K, warning level, trigger time, associated train number, and suggested measures, and is simultaneously pushed to the track maintenance platform, train dispatching system, and GIS platform.

7. The method for graded early warning of vibration anomalies in high-speed railway subgrade based on deep learning according to claim 1, characterized in that, The specific deployment of the accelerometer array is as follows: An accelerometer array is arranged at fixed intervals on both sides of the track shoulder. Each monitoring point is equipped with a triaxial MEMS accelerometer. The triaxial MEMS accelerometer has a range of ±2g, a resolution greater than 50μg, and an operating temperature range of -40℃ to +85℃. Each accelerometer integrates a BeiDou time synchronization module for synchronizing the time across the entire line.

Images